blip-hypermodulator-v1

BLIP HyperModulator

An enhanced hypernetwork that generates LoRA adapters for BLIP image captioning models based on natural language task descriptions.

Overview

This model is a sophisticated hypernetwork that can generate task-specific LoRA (Low-Rank Adaptation) weights for BLIP image captioning models. Given a natural language description of a desired captioning style or task, it generates appropriate LoRA adapters to modify the BLIP model's behavior.

Features

• Cross-attention: Attention mechanism between task descriptions and layer conditioning
• Image-aware adaptation: Optional conditioning on image features for context-aware LoRA generation
• Adaptive gating: Learned gating mechanisms to control LoRA application strength
• Parameter scaling: Dynamic scaling of LoRA parameters
• Caching: Efficient caching of similar task embeddings
• Task interpolation: Ability to blend different captioning styles

Architecture

• Total Parameters: 515,644,222
• LoRA Rank: 16
• LoRA Alpha: 32
• Task Embedding Size: 384
• Latent Size: 256
• Target Modules: dense, fc1, fc2, intermediate, key, output, projection, qkv, query, value

Usage

from transformers import BlipForConditionalGeneration, BlipProcessor
from blip_hyper_modulator import BlipHyperModulator
from sentence_transformers import SentenceTransformer

# Load base BLIP model
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large")
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")

# Load hypermodulator
hypermodulator = BlipHyperModulator.from_pretrained("path/to/model", blip_model)

# Load text encoder
text_encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

# Generate LoRA weights
lora_weights = hypermodulator.generate_lora_weights(
    task_description="Describe this image like a pirate would",
    text_encoder=text_encoder,
    text_tokenizer=None
)

Model Details

• Base Model: Salesforce/blip-image-captioning-large
• Created: 2025-08-29
• Version: 1.0.0
• Framework: PyTorch
• License: MIT

Citation

If you use this model, please cite:

@misc{blip-hypermodulator,
  title={BLIP HyperModulator: Dynamic LoRA Generation for Image Captioning},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/TanVir17Niloy/blip-hypermodulator-v1}
}